CN109360191B - Image significance detection method based on variational self-encoder - Google Patents

Image significance detection method based on variational self-encoder Download PDF

Info

Publication number
CN109360191B
CN109360191B CN201811113241.1A CN201811113241A CN109360191B CN 109360191 B CN109360191 B CN 109360191B CN 201811113241 A CN201811113241 A CN 201811113241A CN 109360191 B CN109360191 B CN 109360191B
Authority
CN
China
Prior art keywords
image
pixel
encoder
super
super pixel
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201811113241.1A
Other languages
Chinese (zh)
Other versions
CN109360191A (en
Inventor
孙正兴
徐峻峰
李博
胡佳高
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing University
Original Assignee
Nanjing University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Nanjing University filed Critical Nanjing University
Priority to CN201811113241.1A priority Critical patent/CN109360191B/en
Publication of CN109360191A publication Critical patent/CN109360191A/en
Application granted granted Critical
Publication of CN109360191B publication Critical patent/CN109360191B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/90—Determination of colour characteristics
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • General Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Molecular Biology (AREA)
  • Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Probability & Statistics with Applications (AREA)
  • Evolutionary Biology (AREA)
  • Quality & Reliability (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses an image significance detection method based on a variational self-encoder, which comprises the following steps: firstly, an input image is segmented into superpixels, the boundary connectivity of each superpixel is calculated, and a background sample set is screened out according to connectivity values; then constructing a variable molecular encoder depth network model, and training the network by using the obtained background sample data in a random gradient descent mode; and finally, reconstructing all superpixels of the whole image by using the trained variational self-encoder network, and calculating reconstruction errors to obtain a significance value.

Description

Image significance detection method based on variational self-encoder
Technical Field
The invention belongs to the technical field of image processing, and relates to an image saliency detection method based on a variational self-encoder.
Background
Recently, saliency detection, which aims to find important parts in images, has become a hot problem in research in the field of computer vision. The saliency detection is an important ring of visual media processing, and provides effective help for many visual applications such as object segmentation, image retrieval, image content editing and the like.
Generally, in the field of significance detection, existing methods can be classified into top-down or bottom-up. The top-down approach is task-driven and requires supervised training with manually labeled truth images. To better discriminate salient objects from the background, we need to apply advanced information and surveillance methods to improve the accuracy of the saliency map. In contrast, bottom-up approaches typically construct saliency maps using low-level cues, such as features, colors, and spatial distances. One of the most applied principles is the contrast-first principle. This principle is mainly to calculate the color contrast and spatial contrast of a region and the surrounding environment to obtain the saliency value of the region. However, the saliency value of an image region is obtained by calculating the contrast of features in RGB and LAB color spaces of the image region according to the method for detecting a saliency region (201510297326. X). When a complex image is encountered, the contrast of the low-layer features cannot reflect a significant difference, resulting in a poor detection result. In addition, there are also methods based on edge precedence. They assume that regions at the edges of the image are more likely to be background, as in the patent "a saliency measure of an image object" (201711124512.9), where a saliency value of an arbitrary object is calculated by calculating the matching difference of a boundary image block and the nearest object image block. Indeed, there is a great possibility that the edge of the image becomes a background, which is described in document 1: borji, D.N.Sihite, and L.Itti.Salient object detection, Abenchmark.In Computer Vision- -ECCV 2012, pages 414- -429.Springer,2012.1, all of which are identified herein. However, as with most previous methods, it is not reasonable to classify all points on the edge of an image as background points. If the target object appears at the edge of the image, the seed chosen as the background will be inaccurate and directly result in false salient detection results. Meanwhile, when the existing method is used for discovering the mode of the background seed point, the generalization expression capability of the background is poor due to the limitation of the model, and the higher significance of part of the background area is detected.
As can be seen from the research and application in the field of image saliency, the current single image saliency detection method has the following disadvantages:
(1) and only low-level image features are used for saliency calculation, so that the saliency detection result has errors on the semantic level of the image.
(2) When edge priority is utilized, a target object is selected as a background seed point by mistake, resulting in a false significant detection result. The pattern recognition model used is deficient in generalization expression capability, resulting in false detection of background regions as salient objects.
Disclosure of Invention
The purpose of the invention is as follows: the invention aims to solve the technical problem of providing an image saliency detection algorithm based on a variational self-encoder for supporting effective detection of a salient object in an image aiming at the defects of the prior art.
In order to solve the technical problem, the invention discloses an image significance detection algorithm based on a variational self-encoder, which comprises the following steps:
the method comprises the following steps:
step 1, inputting an image and screening a background sample;
step 2, training a variation self-encoder depth network model through a background sample;
and 3, calculating a superpixel reconstruction error through a variational self-encoder depth network model to obtain a significance value of the image.
The step 1 comprises the following steps:
step 1-1, dividing the whole input image into N by using SLIC (simple Linear Iterative Cluster) simple Linear Iterative clustering method1(generally taking a value of 300) super-pixels, and calculating an average CIELab color statistical characteristic value of each super-pixel;
step 1-2, clustering super-pixel CIELab color statistical characteristic values by using a K-means clustering algorithm to obtain image areas, and calculating the boundary connectivity of each image area;
step 1-3, calculating the boundary connectivity of the superpixels in the region;
and 1-4, calculating the probability of the super pixel belonging to the background according to the boundary connectivity of the super pixel.
In step 1-2, the boundary connectivity of each image region is calculated using the following formula:
Figure GDA0002442569100000021
wherein RGmRepresenting the mth area in the picture, bdconn (RG)m) Representing the boundary connectivity of the m-th area, SPiRepresenting the ith superpixel in the image, and BD representing the border area, i.e. the area composed of the outermost superpixels of the image.
In step 1-3, the boundary connectivity of the superpixels in the image region is calculated using the following formula:
bdconn(SPi)＝bdconn(RGm),SPi∈RGm
wherein, bdconn (SP)i) Representing the boundary connectivity of the ith super pixel, the boundary connectivity of the ith super pixel is the same as the boundary connectivity of the image area where the ith super pixel is located,
in the steps 1-4, the probability that the super-pixel belongs to the background is calculated by the following formula:
Figure GDA0002442569100000031
wherein bgPro (SP)i) Representing the probability that the ith super pixel belongs to the background,
Figure GDA0002442569100000032
representing a balance weight, selecting a probability greater than or equal to N that belongs to the background2The (typically 0.8) superpixels constitute the background sample set B.
The step 2 comprises the following steps of,
step 2-1, constructing a variation self-encoder depth network model with the depth of 5, wherein the variation self-encoder depth network model comprises an input layer, a hidden layer I, a variation layer, a hidden layer III and an output layer, two adjacent layers are connected in a full connection mode, and the input layer is connected with the output layerNumber of cells N3(generally 400) and the number of hidden layer elements is N corresponding to the RGB values of the pixels contained in the super-pixel4(generally taking the value as 300), the network is in a symmetrical design form, namely the network structures of an encoder part and a decoder part in the variational self-encoder are symmetrical and consistent;
step 2-2, training a variational self-coder depth network model by using a back propagation and random gradient descent method: the parameter to be trained for each hidden layer is WjAnd bjThe input is a vector x which represents a superpixel SP in a background sample set BiMiddle pixel RGB value, output as vector
Figure GDA0002442569100000033
When the reverse propagation is performed, the input is vector y and the output is vector
Figure GDA0002442569100000034
A non-linear activation function used in the network; in the training process, a loss function L is defined as the sum of the cross entropy of encoding and decoding and KL divergence, and the formula is as follows:
Figure GDA0002442569100000035
where q (z | x) is the Gaussian distribution of the input vector x in the variational space z, and p (x | z) is the output vector
Figure GDA0002442569100000036
Gaussian distribution in variational space z, loss function antecedent
Figure GDA0002442569100000037
Representing the cross-entropy loss, the consequent DKL(q (z | x) | p (z)) is KL (Kullback-Leibler Divergence), and for each hidden layer, a random gradient descent method is respectively used to minimize L to obtain a parameter WjAnd bj。
The step 3 comprises the following steps of,
step 3-1, for the ith superpixel SP in the imagei，xiShow itRGB pixel value, obtaining x through training obtained variational self-coder depth network modeliCorresponding coding and decoding results
Figure GDA0002442569100000038
Thereby calculating the super pixel SPiVariation reconstruction error of
Figure GDA0002442569100000039
Step 3-2, passing the ith super pixel SPiVariation reconstruction error of
Figure GDA00024425691000000310
Calculating to obtain the ith super pixel SPiSignificance value of
Figure GDA00024425691000000311
Step 3-3, obtaining N by adopting the method from step 3-1 to step 3-21And obtaining the significance value of the image according to the significance value of the super pixel.
In step 3-1, the ith super pixel SP is calculated by the following formulaiVariation reconstruction error of
Figure GDA0002442569100000041
Figure GDA0002442569100000042
In step 3-2, the ith super pixel SP is calculated by the following formulaiSignificance value of
Figure GDA0002442569100000043
Figure GDA0002442569100000044
According to the method, the situation that the generalization capability of the model is insufficient in the modeling of the prior background mode is overcome by introducing the variation self-encoder depth network to try to calculate the image significance from the depth space; meanwhile, the selection process of the edge background seed points is optimized by utilizing the boundary connectivity with the superpixel, the possibility that the target object is mistakenly selected as the background sample is greatly reduced, and the accuracy of significance detection is improved. Therefore, the invention has higher application value.
Has the advantages that: the invention has the following advantages: firstly, the method screens background samples based on boundary connectivity detection, obtains a pure background sample set with less errors, and improves the modeling precision of an image non-significant background area; secondly, the method constructs a new variational self-encoder depth network and efficiently trains the variational self-encoder depth network in a random gradient descent mode, so that the remarkable information of the image is explored in a depth space, and the generalization capability of background modeling is improved; finally, the invention provides a graph saliency method based on variational reconstruction error calculation, which can obtain a relatively accurate saliency detection result of an input image and improve the detection precision.
Drawings
The foregoing and other advantages of the invention will become more apparent from the following detailed description of the invention when taken in conjunction with the accompanying drawings.
FIG. 1 is a schematic process flow diagram of the present invention.
Fig. 2 is a schematic diagram of a variational self-encoder structure.
Fig. 3a is a schematic diagram comparing the results of saliency detection on a part of livestock images with the results of detection of other saliency methods.
Fig. 3b is a schematic diagram comparing the result of saliency detection on a part of face image with the detection result of other saliency methods.
Detailed Description
The invention is further explained below with reference to the drawings and the embodiments.
As shown in fig. 1, the present invention discloses an image saliency detection algorithm based on a variational self-encoder, which comprises the following steps:
step 1, screening a background sample.
Step 1.1, using document 2: achanta, A.Shaji, K.Smith, A.Lucchi, P.Fua, and S.S.major. Slic superpixels compounded to state-of-the-art superpixel methods.T-PAMI,34(11):2274 and 2282,2012, SLIC (simple Linear iterative clustering) method divides the whole input image into 300 superpixels and calculates the average CIELab color statistical characteristic value of each superpixel.
Step 1.2, using reference 3: zhu, S.Liang, Y.Wei, and J.Sun, "promotion from robust background detection," in Proceedings of the IEEEcomputer Society Conference on Computer Vision and Pattern Recognition,2014, pp.2814-2821, calculating the probability that a superpixel belongs to the background using the following formula:
Figure GDA0002442569100000051
wherein RGmRepresenting the mth area in the picture, bdconn (RG)m) Representing the boundary connectivity of the m-th area, SPiRepresenting the ith superpixel in the image and BD the bounding area. The boundary connectivity of the superpixels in the region can then be calculated using the following formula:
bdconn(SPi)＝bdconn(RGm),SPi∈RGm
wherein the boundary connectivity of the super-pixel i is the same as the boundary connectivity of the region in which it is located.
Then, the probability of the super pixel belonging to the background can be calculated according to the boundary connectivity of the super pixel, and the calculation formula is as follows:
Figure GDA0002442569100000052
wherein bgPro (SP)i) Representing the background probability of the ith super pixel,
Figure GDA0002442569100000053
belonging to the balance weight. Selecting background probability greater than or equal toA super-pixel equal to 0.8 constitutes the background sample set B.
As shown in fig. 2, for step 2 of the present invention, the training of the variational self-encoder specifically includes:
step 2.1, constructing a variational self-encoder model: a variation self-encoder depth network model with the depth of 5 is constructed, and the variation self-encoder depth network model comprises an input layer, a hidden layer I, a variation layer, a hidden layer III and an output layer, wherein the adjacent two layers are connected in a full connection mode. Where the number of input layer elements is 400, corresponding to the pixel RGB values contained in the super-pixel. The number of hidden layer units is 300, and the network is in a symmetrical design form.
Step 2.2, training a neural network model by using a back propagation and random gradient descent method: use document 4: hinton G E, Salakhutdinov R.reducing the dimensional of data with neural networks [ J ]]504-. The parameter to be trained for each hidden layer is WjAnd bjThe input is a vector x which represents a superpixel SP in a background sample set BiMiddle pixel RGB value, output as vector
Figure GDA0002442569100000061
When the reverse propagation is performed, the input is vector y and the output is vector
Figure GDA0002442569100000062
Figure GDA0002442569100000063
Is a non-linear activation function used in the network; during the training process, the loss function is defined as the sum of the encoding-decoding cross entropy and the KL divergence, and the formula is as follows:
Figure GDA0002442569100000064
where q (z | x) is the Gaussian distribution of the input vector x in the variational space z, and p (x | z) is the output vector
Figure GDA0002442569100000068
In changeThe gaussian distribution in the subspace z is passed, the front term of the loss function is the cross entropy, and the back term is the KL divergence. For each hidden layer, a random gradient descent method is respectively used, L is minimized, and a parameter W is obtainedjAnd bj；
And 3, calculating a super-pixel reconstruction error to obtain a significance value.
For all superpixels SP in the imagei，xiRepresenting its RGB pixel values. Training derived variational self-coder depth network for each xiCan obtain the corresponding encoding-decoding result
Figure GDA0002442569100000065
Then the metameric reconstruction error for the superpixel can be computed:
Figure GDA0002442569100000066
the saliency value of the super-pixel can then be found by the following formula:
Figure GDA0002442569100000067
using document 5: dingwen Zhang, Deyu Meng, and Junwei Han.2017.Co-saliencydetection via a self-contained multiple-interaction learning frame. IEEEtransactions on pattern analysis and machine interaction 39,5(2017), 865-.
Use document 6: dingwen Zhang, Junweii Han, Chao Li, Jingdong Wang, and Xuelong Li.2016.detection of Co-medical Objects by Looking Deep and Wide International Journal of Computer Vision 120,2(2016), 215-232. https:// doi. org/10.1007/s11263-016-
Use document 7: the results of the publications such as Xiaochun Cao, Zhijiang Tao, Bao Zhuang, Huazhu Fu, and WeiFeng.2014.Selfadaptive weighted co-preservation video transmission and Processing 23,9(2014), 4175-4186. https:// doi. org/10.1109/TIP.2014. are comparative results to further illustrate the advantages of the effects of the present invention.
FIG. 3a is a schematic diagram showing the comparison of the results of saliency detection of a portion of an animal image with the results of other saliency detection methods; in fig. 3a, the first line is the original image, the second line is the true value image, the third line is the result of the present invention, the fourth line is the result of document 5, the fifth line is the result of document 6, and the sixth line is the result of document 7, so that the effect is seen, compared with the existing significant detection method, the significant detection result of the present invention can better mark the complete contour of the significant object (livestock), the distribution of the significant values in the significant object is more balanced, and simultaneously the significant values of the background area (grassland, shrub, forest, etc.) can be better suppressed.
FIG. 3b is a schematic diagram comparing the result of saliency detection on a part of face image with the detection results of other saliency methods; in fig. 3b, the first line is the original image, the second line is the true value image, the third line is the result of the present invention, the fourth line is the result of document 5, the fifth line is the result of document 6, and the sixth line is the result of document 7, so that the effect is seen, compared with the existing significant detection method, the significant detection result of the present invention can better mark the complete contour of the significant object (human face), the distribution of the significant values in the significant object is more balanced, the significant values in the background area can be better suppressed, and the interference of the non-significant area (clothes and neck under the human face) is reduced.
The present invention provides a method for detecting image saliency based on a variational self-encoder, and a plurality of methods and approaches for implementing the technical solution, the above description is only a preferred embodiment of the present invention, and it should be noted that, for those skilled in the art, a plurality of improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be regarded as the protection scope of the present invention. All the components not specified in the present embodiment can be realized by the prior art.

Claims (4)

1. An image saliency detection method based on a variational self-encoder is characterized by comprising the following steps:
step 1, inputting an image and screening a background sample;
step 2, training a variation self-encoder depth network model through a background sample;
step 3, calculating a superpixel reconstruction error through a variational self-encoder depth network model to obtain a significance value of the image;
the step 1 comprises the following steps:
step 1-1, dividing the whole input image into N by using SLIC simple linear iterative clustering method1Calculating the average CIELab color statistical characteristic value of each super pixel;
step 1-2, clustering super-pixel CIELab color statistical characteristic values by using a K-means clustering algorithm to obtain image areas, and calculating the boundary connectivity of each image area;
step 1-3, calculating the boundary connectivity of the superpixels in the image area;
1-4, calculating the probability of the super pixel belonging to the background according to the boundary connectivity of the super pixel;
in step 1-2, the boundary connectivity of each image region is calculated using the following formula:
Figure FDA0002442569090000011
wherein RGmRepresenting the mth area in the picture, bdconn (RG)m) Representing the boundary connectivity of the m-th area, SPiRepresenting the ith super pixel in the image, and BD represents a boundary area, namely an area formed by the super pixels at the outermost periphery of the image;
in step 1-3, the boundary connectivity of the superpixels in the image region is calculated using the following formula:
bdconn(SPi)＝bdconn(RGm)，SPi∈RGm，
wherein, bdconn (SP)i) The boundary connectivity of the ith super pixel is represented, and the boundary connectivity of the ith super pixel is the same as the boundary connectivity of the image area where the ith super pixel is located;
in the steps 1-4, the probability that the super-pixel belongs to the background is calculated by the following formula:
Figure FDA0002442569090000012
wherein bgPro (SP)i) Representing the probability that the ith super pixel belongs to the background,
Figure FDA0002442569090000013
representing a balance weight, selecting a probability greater than or equal to N that belongs to the background2The super pixels form a background sample set B;
the step 2 comprises the following steps of,
step 2-1, constructing a variation self-encoder depth network model with the depth of 5, wherein the variation self-encoder depth network model comprises an input layer, a hidden layer I, a variation layer, a hidden layer III and an output layer, the adjacent two layers are connected in a full connection mode, and the number of input layer units is N3The number of hidden layer units is N corresponding to the RGB values of the pixels contained in the super-pixel4The network is in a symmetrical design form, namely the network structures of an encoder part and a decoder part in the variational self-encoder are symmetrical and consistent;
step 2-2, training a variational self-coder depth network model by using a back propagation and random gradient descent method: the parameter to be trained for each hidden layer is WjAnd bjThe input is a vector x which represents a superpixel SP in a background sample set BiMiddle pixel RGB value, output as vector
Figure FDA0002442569090000021
When the reverse propagation is performed, the input is vector y and the output is vector
Figure FDA0002442569090000022
Figure FDA0002442569090000023
Is a non-linear activation function used in the network; in the training process, a loss function L is defined as the sum of the cross entropy of encoding and decoding and KL divergence, and the formula is as follows:
Figure FDA0002442569090000024
where q (z | x) is the Gaussian distribution of the input vector x in the variational space z, and p (x | z) is the output vector
Figure FDA0002442569090000025
Gaussian distribution in variational space z, loss function antecedent
Figure FDA0002442569090000026
Representing the cross-entropy loss, the consequent DKL(q (z | x) | p (z)) is KL divergence, and for each hidden layer, a random gradient descent method is respectively used to minimize L to obtain a parameter WjAnd bj。
2. The method of claim 1, wherein step 3 comprises the steps of,
step 3-1, for the ith superpixel SP in the imagei，xiRepresenting RGB pixel values, obtaining x through a variation self-coder depth network model obtained through trainingiCorresponding coding and decoding results
Figure FDA0002442569090000027
Thereby calculating the super pixel SPiVariation reconstruction error of
Figure FDA0002442569090000028
Step 3-2, passing the ith super pixel SPiVariation reconstruction error of
Figure FDA0002442569090000029
Calculating to obtain the ith super pixel SPiSignificance value of
Figure FDA00024425690900000210
Step 3-3, extractingObtaining N by the method of step 3-1 to step 3-21And obtaining the significance value of the image according to the significance value of the super pixel.
3. The method according to claim 2, wherein in step 3-1, the ith super pixel SP is calculated by the following formulaiVariation reconstruction error of
Figure FDA00024425690900000211
Figure FDA00024425690900000212
4. The method according to claim 3, wherein in step 3-2, the ith super pixel SP is calculated by the following formulaiSignificance value of
Figure FDA00024425690900000213
Figure FDA00024425690900000214
CN201811113241.1A 2018-09-25 2018-09-25 Image significance detection method based on variational self-encoder Active CN109360191B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201811113241.1A CN109360191B (en) 2018-09-25 2018-09-25 Image significance detection method based on variational self-encoder

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201811113241.1A CN109360191B (en) 2018-09-25 2018-09-25 Image significance detection method based on variational self-encoder

Publications (2)

Publication Number Publication Date
CN109360191A CN109360191A (en) 2019-02-19
CN109360191B true CN109360191B (en) 2020-06-12

Family

ID=65351366

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201811113241.1A Active CN109360191B (en) 2018-09-25 2018-09-25 Image significance detection method based on variational self-encoder

Country Status (1)

Country Link
CN (1) CN109360191B (en)

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109993208B (en) * 2019-03-04 2020-11-17 北京工业大学 Clustering processing method for noisy images
JP7183904B2 (en) * 2019-03-26 2022-12-06 日本電信電話株式会社 Evaluation device, evaluation method, and evaluation program
CN110517759B (en) * 2019-08-29 2022-03-25 腾讯医疗健康(深圳)有限公司 Method for determining image to be marked, method and device for model training
CN114677368B (en) * 2022-04-19 2022-10-25 中国人民解放军32021部队 Image significance detection method and device

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106529419A (en) * 2016-10-20 2017-03-22 北京航空航天大学 Automatic detection method for significant stack type polymerization object in video
CN108171701A (en) * 2018-01-15 2018-06-15 复旦大学 Conspicuousness detection method based on U networks and confrontation study

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9870503B2 (en) * 2015-12-31 2018-01-16 International Business Machines Corporation Visual object and event detection and prediction system using saccades

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106529419A (en) * 2016-10-20 2017-03-22 北京航空航天大学 Automatic detection method for significant stack type polymerization object in video
CN108171701A (en) * 2018-01-15 2018-06-15 复旦大学 Conspicuousness detection method based on U networks and confrontation study

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
Background Prior-Based Salient Object Detection via Deep Reconstruction Residual;Junwei Han et.al;《IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY》;20150831;第25卷(第8期);第1311-1313页 *
Co-saliency Detection via Sparse reconstruction and Co-salient Object Discovery;Bo Li et.al;《PCM 2017: Advances in Multimedia Information Processing》;20180510;第224-225页 *
Saliency Optimization from Robust Background Detection;Wangjiang Zhu et.al;《CVPR2014》;20141231;第2-4页 *
Subitizing with Variational Autoencoders;Rijnder Wever et.al;《arXiv:1808.00257v1 [cs.CV]》;20180801;第3-5页 *

Also Published As

Publication number Publication date
CN109360191A (en) 2019-02-19

Similar Documents

Publication Publication Date Title
Zhang et al. Triplet-based semantic relation learning for aerial remote sensing image change detection
CN109360191B (en) Image significance detection method based on variational self-encoder
CN110428432B (en) Deep neural network algorithm for automatically segmenting colon gland image
CN109241913B (en) Ship detection method and system combining significance detection and deep learning
US9042648B2 (en) Salient object segmentation
CN106897681B (en) Remote sensing image contrast analysis method and system
Cao et al. A new change-detection method in high-resolution remote sensing images based on a conditional random field model
CN108960404B (en) Image-based crowd counting method and device
Cao et al. Automatic change detection in high-resolution remote-sensing images by means of level set evolution and support vector machine classification
CN105122308A (en) Systems and methods for multiplexed biomarker quantitation using single cell segmentation on sequentially stained tissue
CN107784663A (en) Correlation filtering tracking and device based on depth information
CN108629286B (en) Remote sensing airport target detection method based on subjective perception significance model
US20220215548A1 (en) Method and device for identifying abnormal cell in to-be-detected sample, and storage medium
WO2016066042A1 (en) Segmentation method for commodity picture and device thereof
Liu et al. A night pavement crack detection method based on image‐to‐image translation
CN113591968A (en) Infrared weak and small target detection method based on asymmetric attention feature fusion
CN106157330B (en) Visual tracking method based on target joint appearance model
Yang et al. Classified road detection from satellite images based on perceptual organization
CN116645592B (en) Crack detection method based on image processing and storage medium
Lizarazo et al. Fuzzy segmentation for object‐based image classification
CN112258525A (en) Image abundance statistics and population recognition algorithm based on bird high frame frequency sequence
CN111881915A (en) Satellite video target intelligent detection method based on multiple prior information constraints
Wang et al. Salient object detection by robust foreground and background seed selection
CN110910497B (en) Method and system for realizing augmented reality map
Pang et al. Salient object detection via effective background prior and novel graph

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant